Inspection apparatus, target variable correction method, and target variable correction program

Abstract

To provide an inspection apparatus, a method for correcting a target variable, and a target variable correction program that improve the inference accuracy of a learning model while correcting mislabeled data, even when arbitrary data are used as training data in inspection using a learning model.SOLUTION: An inspection apparatus according to one embodiment of the present disclosure includes a learning unit, and a teacher data estimation unit having an inference unit, a weighting unit, and a correction unit. The learning unit learns a learning model using a dataset of teacher data including labels as target variables. The inference unit performs inference using the learning model with the dataset of the teacher data. The weighting unit weights, with an arbitrary weighting value, a correction amount indicated by the difference between a first feature corresponding to each teacher data in a feature space of the learning model and a second feature obtained through the inference. The teacher data correction unit corrects the label on the basis of the weighted correction amount.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present disclosure relates to an inspection apparatus, a response variable correction method, and a response variable correction program. [Background technology]

[0002] In the past, a technology has been known for a product inspection process in which data on the product to be inspected is input to a learning model trained by machine learning to infer whether the product is good or bad. The learning model is trained using a data set of training data consisting of explanatory variables and a target variable, so that when an explanatory variable is input, the target variable is output.

[0003] Training data is often manually labeled as a target variable through annotation work. As a result, training data sets can contain training data that has been mislabeled (hereinafter referred to as "mislabeled") due to human error. To address this issue, Patent Document 1 discloses a technology for correcting mislabeled data using a pre-trained learning model.

[0004] In this conventional technology, for a learning model that performs image segmentation processing, the image to be processed is input into a pre-trained learning model, and the output processed image is compared with the training data image used in the pre-training to evaluate whether the label is appropriate. If the label is evaluated to be inappropriate, the label is corrected. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2020 / 031243 Summary of the Invention [Problem to be solved by the invention]

[0006] However, the above-mentioned conventional techniques use image data that is premised on image segmentation processing as training data, and cannot be directly applied when arbitrary data is used as training data.

[0007] The present disclosure provides a technology that can improve the inference accuracy of a learning model while correcting mislabels, even when any data is used as training data in testing using a learning model. [Means for solving the problem]

[0008] An inspection device according to one aspect of the present disclosure includes a learning unit, an inference unit, a weighting unit, and a correction unit. The learning unit learns a learning model using a dataset of training data including a label as a target variable. The inference unit performs inference using the learning model using the dataset of training data. The weighting unit weights, with an arbitrary weight value, a correction amount indicated by a difference between a first feature corresponding to each of the training data in the feature space of the learning model and a second feature obtained by the inference. The correction unit corrects the label based on the weighted correction amount. [Effects of the Invention]

[0009] According to the present disclosure, even when any data is used as training data, it is possible to improve the inference accuracy of a learning model while correcting mislabels. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram (part 1) outlining a method for correcting a response variable according to an embodiment. [Figure 2] FIG. 2 is a schematic explanatory diagram (part 2) of the method for correcting a response variable according to the embodiment. [Figure 3] FIG. 3 is a diagram (part 3) outlining the method for correcting a response variable according to the embodiment. [Figure 4] FIG. 4 is a schematic explanatory diagram (part 4) of the method for correcting a response variable according to the embodiment. [Figure 5] FIG. 5 is a block diagram showing an example of the configuration of an inspection device according to an embodiment. [Figure 6] FIG. 6 is a block diagram illustrating an example of the configuration of the teacher data correction unit. [Figure 7] FIG. 7 is a flowchart showing a processing procedure executed by the inspection device according to the embodiment. [Figure 8] FIG. 8 is a flowchart showing the processing procedure of the teacher data correction processing. [Figure 9] FIG. 9 is a diagram showing an example of a presentation layout presented by the presentation unit. [Figure 10] FIG. 10 is a diagram illustrating an outline of a method for correcting a response variable according to a modified example. [Figure 11] FIG. 11 is a diagram showing an example of image data obtained by visualizing sound. [Figure 12] FIG. 12 is a hardware configuration diagram showing an example of a computer that realizes the functions of the inspection device. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In the following embodiments, the same components are designated by the same reference numerals, and redundant description will be omitted.

[0012] In the following description, the inspection device 10 (see FIG. 5) according to an embodiment of the present disclosure (hereinafter referred to as "the present embodiment") is assumed to be a device that inspects products using the learning model 11b (see FIG. 4). The objective variable correction method according to the present embodiment is assumed to be a objective variable correction method executed by the inspection device 10.

[0013] In this embodiment, the learning model 11b is trained using image data as training data. This image data does not require specific image processing such as segmentation processing. This image data is, for example, image data obtained by visualizing sound.

[0014] In this embodiment, the training data of the learning model 11b is image data, but the training data of the learning model 11b may be data other than image data. This embodiment can also be applied to cases where the training data is, for example, audio data, text data, etc.

[0015] The present disclosure will be described in the following order. 1. Overview 2. Example of inspection equipment configuration 3. Processing Procedure 4. Presentation layout 5. Variations 6. Hardware Configuration 7. Conclusion

[0016] <<1. Overview>> First, an overview of the response variable correction method according to the present embodiment will be described with reference to FIGS. 1 to 4. FIG. 1 is a schematic explanatory diagram (part 1) of the response variable correction method according to the embodiment. FIG. 2 is a schematic explanatory diagram (part 2) of the response variable correction method according to the embodiment. FIG. 3 is a schematic explanatory diagram (part 3) of the response variable correction method according to the embodiment. FIG. 4 is a schematic explanatory diagram (part 4) of the response variable correction method according to the embodiment.

[0017] In existing technologies that use learning models to inspect products, the training data used to train the model is often manually labeled as a target variable. As a result, the training data set (hereinafter referred to as the "training data set") can contain mislabeled training data due to human error.

[0018] Specifically, as shown in Figure 1, real-world training datasets may contain mislabeled training data, where training data Td1 and Td2 are labeled as "NG" even though they should be labeled "OK." Similarly, training data Td3, which should be labeled "NG," may contain mislabeled training data, where it is labeled as "OK."

[0019] If a learning model is trained using this real-world training dataset, there is a risk that unintended erroneous judgments based on mislabeling will occur when the learning model judges whether a product is good or bad after training.

[0020] To address this problem, the method for correcting a response variable according to this embodiment automatically corrects, for example, the mislabeled training data Td1, Td2, and Td3 shown in FIG. 1 to obtain an ideal training data set.

[0021] In brief, in the method for correcting a target variable according to this embodiment, the inspection device 10 trains a learning model 11b using a teacher dataset and performs inference using the learning model 11b, which uses the teacher dataset used for training as input. The inspection device 10 also weights, with an arbitrary weight value, the correction amount indicated by the difference between a first feature corresponding to each piece of teacher data in the feature space of the learning model 11b and a second feature obtained by the inference. The inspection device 10 then corrects the label based on the weighted correction amount. The weight value is the weight in the weighted moving average.

[0022] The method for correcting a target variable according to the embodiment will be described in detail below. First, definitions of terms used in the following description are shown in FIG. 2. As shown in FIG. 2, in the following description, the label of each training data in the training data set of the learning model 11b is expressed as Y m Let's say.

[0023] Also, for each label Y m The feature of y m The feature value y m is an example of a "first feature amount." m is the label Y of each training data in the feature space of the learning model 11b. m The feature value corresponds to the score. Note that the term "feature value" can also be called "score" or "weight."

[0024] In addition, the inference label, which is the label of each training data indicated by the inference result of the above inference, is expressed as Y' m Also, each label Y'm The inferred value corresponding to y' m Let the inferred value y' m corresponds to an example of a "second feature amount."

[0025] The correction amount is denoted by ω and the weight value is denoted by η. The weight value η is the weight of the moving average as described above. The correction amount weighted by the weight value η is denoted by ω. m The variable m corresponds to the number of times the teacher data correction process is executed in the model optimization process executed in the target variable correction method according to this embodiment. m is an arbitrary natural number.

[0026] An image of the method for correcting the objective variable according to this embodiment is shown in Fig. 3. On the right side of Fig. 3, when the inspection device 10 inputs the teacher data set used for learning and performs inference using the learned learning model 11b, the inferred value y' corresponding to each teacher data in the feature space is shown. m The distribution of the inferred value y' is shown schematically. m =0.5 indicates the boundary between the labels "OK" and "NG" in this feature space.

[0027] As in the case of the training data Td0 and Td4 in Figure 3, the respective inferred values ​​y' m The inferred label Y' corresponding to m and the uncorrected label Y m-1 If they match, the label is not mislabeled, so the inspection device 10 sets the correction amount ω to 0 for the teacher data Td0 and Td4, and does not treat them as correction targets.

[0028] On the other hand, as in the case of training data Td2 and Td3, the respective inferred values ​​y' m The inferred label Y' corresponding to m and the uncorrected label Y m-1 If there is no match, the inspection device 10 determines the inferred label Y' m and the uncorrected label Y m-1 Then, the inspection device 10 calculates the correction amount ω indicated by the difference between the pre-correction label Y m-1 After correcting the label of the training data to be corrected, the label Ym Let's say.

[0029] At this time, in the response variable correction method according to the embodiment, the inspection device 10 weights the correction amount ω with an arbitrary weight value.

[0030] Specifically, in the method for correcting a target variable according to the embodiment, the inspection device 10 performs inference using the current learning model 11b with the current teacher data set as input. That is, the inspection device 10 calculates each inferred value y' using the current learning model 11b for all teacher data included in the current teacher data set 11a. m Here, the term "current" may be read as "latest."

[0031] Next, the inspection device 10 calculates the current label of each training data, i.e., the pre-correction label Y m-1 The feature y corresponding to m-1 The calculated inferred value y' m The correction amount ω indicated by the difference between is calculated for all training data.

[0032]

number

[0033] Then, the inspection device 10 weights each correction amount ω by an arbitrary weight value η in the weighted moving average according to the following equation (b): m corresponds to the degree of correction of each training data. m The larger the absolute value of ω, the higher the correction degree. m The smaller the absolute value of η, the lower the degree of correction. The weight value η can be arbitrarily designated by the user. The weight value η satisfies η>0.

[0034]

number

[0035] Then, the inspection apparatus 10 calculates each weighted correction amount ω m Based on this, the label Y of each training data is calculated using the following formulas (c) and (d). m Fix.

[0036]

number

number

[0037] The weight value η is used as a parameter for adjusting the correction amount ω.

[0038] The model optimization process in the method for correcting a target variable according to the embodiment will now be described in more detail. As shown in Fig. 4, the inspection device 10 executes the model optimization process. The model optimization process is a process in which learning using a teacher dataset is repeated while correcting mislabels in the teacher dataset, thereby optimizing the learning model 11b.

[0039] As shown in Fig. 4, in the model optimization process, the inspection device 10 executes a learning process (step S1) and a teacher data correction process (step S2). As shown in Fig. 4, in the learning process, the inspection device 10 repeats learning of the learning model 11b using the teacher data set 11a as input n times for one teacher data correction process. The variable n corresponds to the number of learning times performed for one teacher data correction process in the model optimization process. n is an arbitrary natural number.

[0040] In the model optimization process, the inspection device 10 repeats the teacher data correction process m times. In each teacher data correction process, the inspection device 10 executes inference using the current learning model 11b with the current teacher data set 11a as input (step S21).

[0041] Next, the inspection device 10 calculates the correction amount ω for all the training data using the above formula (a) (step S22). Then, the inspection device 10 weights each correction amount ω with a weight value η using the above formula (b) (step S23).

[0042] Then, the inspection apparatus 10 calculates each weighted correction amount ω m Based on this, the label Y of each training data is calculated using the above formulas (c) and (d). m Fix.

[0043] This makes it possible to improve the inference accuracy of the learning model 11b while correcting mislabels, even when any data is used as training data.

[0044] An example of the configuration of the inspection device 10 to which the method for correcting a response variable according to this embodiment is applied will be described in more detail below.

[0045] <<2. Inspection equipment configuration example>> Fig. 5 is a block diagram showing a configuration example of the inspection device 10 according to an embodiment of the present disclosure. Fig. 6 is a block diagram showing a configuration example of the teacher data correction unit 12c. Note that Figs. 5 and 6 show functional blocks of only components necessary for explaining this embodiment, and general components are omitted.

[0046] In addition, in the description using FIGS. 5 and 6, the description of components that have already been described will be appropriately simplified or omitted.

[0047] 5, the inspection device 10 includes a storage unit 11 and a control unit 12. The inspection device 10 is also connected to an annotation device 3, a UI (User Interface) device 5, and an inspection data acquisition device 7.

[0048] The annotation device 3 is a device used by an annotation worker when performing annotation work. The annotation device 3 has a database that stores, for example, the original data of the training data, and the worker labels each piece of original data stored in this database. The annotation device 3 outputs the labeled dataset after the annotation work as a training dataset.

[0049] The annotation device 3 may be a device that is directly connected to the inspection device 10 via wired or wireless communication, or may be a device that is indirectly connected via a recording medium to which the teacher dataset is output.

[0050] The UI device 5 is a device including an input / output interface for various information related to the model optimization process. The UI device 5 includes an input interface for accepting various inputs related to the model optimization process from a user. The UI device 5 includes, for example, an input interface for accepting input of an arbitrary weight value η from a user. The UI device 5 includes, for example, a keyboard, a mouse, etc. as input interfaces.

[0051] The UI device 5 also includes an output interface that presents various types of information related to the model optimization process to the user. The UI device 5 includes, for example, a display as the output interface. The display is, for example, an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) display.

[0052] The UI device 5 may be configured such that the input interface and the output interface are integrated into one unit, for example, by a touch panel display. In this case, the input interface may include software components such as a GUI (Graphical User Interface).

[0053] The inspection data acquisition device 7 is a device that acquires inspection data of an inspection target. When the inspection data is, for example, a captured image of a product to be inspected, the inspection data acquisition device 7 is configured as an image sensor that captures an image of the product. In this case, the inspection data acquisition device 7 includes, for example, a CMOS (Complementary Metal Oxide Semiconductor).

[0054] Furthermore, if the inspection data is, for example, image data of the sound emitted by the product being inspected, the inspection data acquisition device 7 is configured as a device that collects and images the sound emitted by the product. In this case, the inspection data acquisition device 7 includes, for example, a microphone and a digital processor, and analyzes and images the sound emitted by the product recorded using the microphone. The inspection data acquisition device 7 outputs, for example, a spectrogram that visualizes the sound through frequency analysis.

[0055] In addition to these, the test data may be voice data itself, text data, etc. The test data acquisition device 7 may be configured appropriately depending on the type of test data. The test data acquisition device 7 is placed, for example, on a product production line, etc. The test data acquisition device 7 may be a device that communicates with the test device 10 via wired communication or wireless communication.

[0056] The storage unit 11 is realized by a storage device such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, or an HDD (Hard Disk Drive).

[0057] 5, the storage unit 11 stores a teacher data set 11a, a learning model 11b, and inference result information 11c. The teacher data set 11a stores a group of teacher data used when learning the learning model 11b.

[0058] The learning model 11b is an AI (Artificial Intelligence) model for determining whether a product is good or bad. The learning model 11b is, for example, a DNN (Deep Neural Network) model that is trained using a deep learning algorithm.

[0059] The learning model 11b is read as a DNN model into the inspection unit 12e (described later), and then trained so that when inspection data to be inspected is input to the inspection unit 12e, the inspection unit 12e can determine whether the inspection data is good or bad. That is, the inspection unit 12e that has read the learning model 11b operates as an inspection AI that inspects products.

[0060] The inference result information 11c stores and accumulates the inference results when the teacher data correction unit 12c (described later) performs inference using the learning model 11b with the teacher data set 11a as input in the model optimization process. The inference result information 11c also stores and accumulates the feature quantity y of the nearest label of all the teacher data based on this inference result. m and this feature y m The nearest label Y corresponding to m Includes.

[0061] The control unit 12 corresponds to a so-called processor or controller. The control unit 12 is realized by, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or a GPU (Graphical Processing Unit). The control unit 12 executes a target variable correction program according to an embodiment (not shown) stored in the storage unit 11, using RAM as a work area. The control unit 12 can also be realized by, for example, an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0062] The control unit 12 has an acquisition unit 12a, a learning unit 12b, a teacher data correction unit 12c, a presentation unit 12d, and an inspection unit 12e, and realizes or executes the functions and actions of information processing described below.

[0063] The acquisition unit 12a acquires the teacher dataset after annotation work by the person in charge from the annotation device 3, and stores it as the teacher dataset 11a in the storage unit 11. Note that the teacher dataset 11a immediately after being acquired by the acquisition unit 12a from the annotation device 3 may contain mislabeled teacher data.

[0064] The learning unit 12b executes the learning process of step S1 in the model optimization process described with reference to Fig. 4. The teacher data correction unit 12c executes the teacher data correction process of step S2 in the model optimization process.

[0065] As shown in Fig. 6, the teacher data correction unit 12c includes an inference unit 12ca, a weighting unit 12cb, and a correction unit 12cc. The inference unit 12ca executes step S21 in the model optimization process. Specifically, the inference unit 12ca executes inference using the current learning model 11b with the current teacher data set 11a as input. The inference unit 12ca also reflects the inference result in the inference result information 11c.

[0066] The weighting unit 12cb executes steps S22 and S23 in the model optimization process. Specifically, in step S22, the weighting unit 12cb calculates a correction amount ω for all training data using the above formula (a). In addition, in step S23, the weighting unit 12cb weights each correction amount ω with a weight value η using the above formula (b).

[0067] The correction unit 12cc executes step S24 in the model optimization process. Specifically, the correction unit 12cc calculates each weighted correction amount ω m Based on this, the label Y of each training data is calculated using the above formulas (c) and (d). mThe correction unit 12cc also reflects the correction result in the teacher dataset 11a and the inference result information 11c.

[0068] Returning to the description of Fig. 5, the presentation unit 12d presents various pieces of information relating to at least the model optimization process to the user via the UI device 5. The presentation unit 12d presents, on the UI device 5, a GUI relating to the model optimization process, for example.

[0069] The presentation unit 12d presents to the user, for example, a GUI that allows the user to specify the weight value η. The presentation unit 12d also presents to the user a correction amount ω weighted according to the weight value η. m A specific example of the proposed layout that the presenting unit 12d presents to the UI device 5 will be described later with reference to FIG.

[0070] The inspection unit 12e operates as an inspection AI by reading the learning model 11b optimized by the model optimization process. The inspection unit 12e inputs the inspection data acquired from the inspection data acquisition device 7 into the learning model 11b, acquires the inspection results by the learning model 11b, and causes the presentation unit 12d to present the results, for example.

[0071] <<3. Processing Procedure>> Next, the processing procedure executed by the inspection device 10 will be described with reference to Fig. 7 and Fig. 8. Fig. 7 is a flowchart showing the processing procedure executed by the inspection device 10 according to the embodiment. Fig. 8 is a flowchart showing the processing procedure of the teacher data correction process.

[0072] First, the acquisition unit 12a acquires the teacher data set 11a (step S101). The acquisition unit 12a acquires the teacher data set 11a from the annotation device 3, for example, as described above.

[0073] Next, the learning unit 12b initializes the learning model 11b prior to the learning process (step S102). As an example of initialization, the learning unit 12b sets default values ​​to the parameters of the learning model 11b. As another example of initialization, the learning unit 12b sets hyperparameters that specify "how to perform learning" for the learning model 11b. Additionally, the learning unit 12b sets an initial value of 1 to a counter m for the number of loops of the model optimization process and a counter n for the number of learning times (step S103).

[0074] Then, the learning unit 12b performs learning using the current teacher data set 11a. That is, the learning unit 12b updates the model parameters of the learning model 11b with the current teacher data set 11a (step S104).

[0075] After executing step S104, the learning unit 12b increments the counter n (step S105) and determines whether the counter n has exceeded a predetermined value (step S106). The "predetermined value" here refers to the upper limit threshold for the number of learning operations to be performed prior to one training data correction process.

[0076] If the counter n does not exceed the predetermined value (step S106, No), the learning unit 12b repeats the process from step S104. If the counter n exceeds the predetermined value (step S106, Yes), the teacher data correction unit 12c executes the teacher data correction process (step S107).

[0077] 8, in the teacher data correction process, the inference unit 12ca executes inference using the current learning model 11b with the current teacher data set 11a as input (step S201). Specifically, the inference unit 12ca calculates each inferred value y' using the current learning model 11b for all teacher data included in the current teacher data set 11a. m Calculate.

[0078] Then, the weighting unit 12cb calculates the feature value y corresponding to the current label of each training data by the above formula (a). m-1and the inferred value y' calculated in step S201. m The correction amount ω indicated by the difference between m and m is calculated for all training data (step S202). Note that the feature amount y0 when m=1 is the same as the label Y0 immediately after annotation.

[0079] Next, the weighting unit 12cb weights each correction amount ω calculated in step S202 by an arbitrary weight value η in the weighted moving average using the above formula (b) (step S203).

[0080] Then, the correction unit 12cc calculates each weighted correction amount ω m Based on this, the label Y of each training data is calculated using the above formulas (c) and (d). m (step S204), and the teacher data correction process ends.

[0081] Returning to the explanation of Figure 7, after executing the teacher data correction process in step S107, the teacher data correction unit 12c increments counter m and sets counter n to its initial value of 1 (step S108). Then, the teacher data correction unit 12c determines whether counter m has exceeded a predetermined value (step S109). The "predetermined value" here refers to the upper threshold of the number of model optimization loops.

[0082] If the counter m does not exceed the predetermined value (step S109, No), the learning unit 12b repeats the process from step S104. If the counter m exceeds the predetermined value (step S109, Yes), the model optimization process ends.

[0083] <<4. Presentation Layout>> Next, an example of a presentation layout presented by the presentation unit 12d will be described. Fig. 9 is a diagram showing an example of a presentation layout presented by the presentation unit 12d. The presentation unit 12d presents to the UI device 5, for example, a presentation layout as shown in Fig. 9.

[0084] 9, the presentation layout includes, for example, operation components G1 to G5, a correction candidate display section 51, and an inference accuracy display section 52. The operation component G1 is a GUI that allows the user to select a teacher data file corresponding to the teacher dataset 11a.

[0085] The operation item G2 is a GUI that allows the user to select a file in which to save the corrected teacher dataset 11a. The operation item G3 is a GUI that allows the user to arbitrarily specify the weight value η in the model optimization process. Note that the presentation unit 12d may present, for example, a recommended value for the weight value η before the user specifies the weight value η. In this case, the recommended value may be, for example, the weight value η specified in a past model optimization process, or the weight value η that has the greatest track record of improving the accuracy rate, i.e., the inference accuracy, of the learning model 11b among the weight values ​​η specified in the past.

[0086] When a new weight value η is specified for the operation element G3 by the user, the operation element G4 is configured to assign a correction amount ω m This is a GUI that recalculates the

[0087] The correction candidate display unit 51 displays the correction amount ω weighted by the weight value η specified for the operation item G3. m If the training data is image data, the presentation unit 12d presents a group of images that are candidates for correction on the correction candidate display unit 51 as shown in FIG.

[0088] The presentation unit 12d also presents, together with each correction candidate, the label names before and after the correction for each correction candidate and an operation component G5, which is a GUI for selecting each correction candidate, on the correction candidate display unit 51. The user can exclude a correction candidate for which the user does not want to correct the label by deselecting it using the operation component G5. Figure 9 shows that image #3 has been excluded from the correction candidates.

[0089] The operation item G6 is a GUI that executes correction by the correction unit 12cc targeting the correction candidate selected in the correction candidate display unit 51, and saves the corrected teacher dataset 11a.

[0090] The inference accuracy display unit 52 displays the inference accuracy before and after the correction of the teacher dataset 11a. The display on the inference accuracy display unit 52 is performed as follows. That is, the learning unit 12b learns the learning model 11b using the teacher dataset 11a corrected by the correction unit 12cc. The inference unit 12ca performs inference using this trained learning model 11b, using the teacher dataset 11a corrected by the correction unit 12cc as input. The presentation unit 12d presents to the user the inference results of the inference unit 12ca before and after the correction by the correction unit 12cc. This inference result includes the accuracy rate of the learning model 11b before and after the correction of the teacher dataset 11a, i.e., the inference accuracy, and the presentation unit 12d displays this inference accuracy on the inference accuracy display unit 52.

[0091] In other embodiments, the presenting unit 12d may be omitted. For example, in other embodiments, the labels of the training data are corrected without prompting the user for confirmation.

[0092] <<5. Modifications>> Although the embodiment of the present disclosure has been described so far, the method for correcting a response variable according to the present embodiment can be modified in several other ways.

[0093] For example, in the above-described embodiment, a case was described in which the error label "NG" in the training data was corrected to "OK" and the error label "OK" was corrected to "NG" for the learning model 11b for determining whether a product is good or bad.

[0094] Here, when determining whether a product is good or bad, it may be acceptable if the product is determined to be defective and is not shipped, for example. However, it is problematic if a defective product is determined to be good and is mistakenly shipped.

[0095] Therefore, the weight value η may be changed according to the inferred label Y'm so as to allow overdetection of defective products, which is when good products are judged as defective products, and prevent underdetection of defective products, which is when bad products are judged as good products. In other words, the inspection device 10 may determine whether or not to correct the labels of the training data according to the weight value η.

[0096] 10 is a diagram illustrating an outline of a target variable correction method according to a modified example. As shown in FIG. 10, in the target variable correction method according to the modified example, the inspection device 10 calculates the inferred label Y' so as to allow correction of the teacher data td1 and Td2 with the error label "NG" and to suppress correction of the teacher data Td3 with the error label "OK". m In order to achieve this, the inspection device 10 uses, for example, the following equation (e).

[0097]

number

[0098] When the above formula (e) is used, the inspection device 10 presents the operation item G3 so that the user can specify two weight values ​​η0 and η1 in the presentation layout shown in Fig. 9. The two weight values ​​are at least different values, as shown by η0>η1 in formula (e).

[0099] In this case, the inspection device 10 calculates a weighted correction amount ω based on each of the two specified weight values ​​η0 and η1. m Inferred label Y' m Then, the inspection device 10 calculates the inferred label Y' as shown in equation (e). m On the other hand, the inspection device 10 uses the weight value η1 that is 1, i.e., "NG" as the weight value η. m The weight value η0 for which is 0, i.e., "OK", is not adopted as the weight value η.

[0100] This makes it possible to prevent a situation in which a defective product is mistakenly determined to be a non-defective product and shipped, for example.

[0101] In the above-described embodiment, the teacher data is image data that visualizes sound. An example of this image data is shown in Fig. 11. Fig. 11 is a diagram showing an example of image data that visualizes sound.

[0102] In this embodiment, as described above, the correction amount ω is arbitrarily weighted by the weight value η in the weighted moving average. That is, in this embodiment, the user can control the degree of correction by arbitrarily adjusting the weight value η. Specifically, the larger the weight value η, the higher the degree of correction, and the smaller the weight value η, the lower the degree of correction.

[0103] For example, if a user wants to correct only so-called marginal values ​​near the boundary between the labels "OK" and "NG" in the feature space, the user can obtain the desired result by decreasing the weight value η. Also, if the user wants to correct values ​​far from the boundary as well, the user can obtain the desired result by increasing the weight value η. Furthermore, as described above, the training data may be data other than image data. That is, according to the method for correcting a target variable according to the embodiment, even when any data is used as training data, it is possible to improve the inference accuracy of the learning model 11b while correcting mislabels.

[0104] In addition, when the training data is image data that visualizes sound, as in this embodiment, the learning model 11b trained using this training data can be used to perform noise inspections of products that incorporate motors, for example.

[0105] In the above-described embodiment, the model optimization process is terminated when the teacher data correction process is repeated a predetermined number of times, but the model optimization process may be terminated based on the inference accuracy of the learning model 11b. In this case, the inspection device 10 may terminate the model optimization process when, for example, the inference accuracy of the learning model 11b exceeds a predetermined inference accuracy.

[0106] Furthermore, among the processes described in the above-described embodiments of the present disclosure, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using known methods. In addition, the information including the processing procedures, specific names, various data, and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified. For example, the various information shown in each drawing is not limited to the information shown in the drawings.

[0107] Furthermore, the components of each device shown in the figure are conceptual functional components and do not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc.

[0108] For example, a learning device specialized for learning and optimizing the learning model 11b can be configured by excluding the test data acquisition device 7 and the test unit 12e from the components shown in Fig. 5. Furthermore, the learning device referred to here or the test device 10 shown in Fig. 5 may be configured integrally with the annotation device 3.

[0109] The above-described embodiments of the present disclosure can be combined as appropriate within the scope of the present disclosure without causing any contradiction in the processing content. The order of the steps shown in the sequence diagrams or flowcharts of the present embodiments can be changed as appropriate.

[0110] <<6. Hardware Configuration>> The inspection device 10 according to the embodiment of the present disclosure described above is realized by a computer 1000 having a configuration as shown in Fig. 12, for example. Fig. 12 is a hardware configuration diagram showing an example of the computer 1000 that realizes the functions of the inspection device 10. The computer 1000 has a CPU 1100, a RAM 1200, a ROM 1300, a secondary storage device 1400, a communication interface 1500, and an input / output interface 1600. The components of the computer 1000 are connected by a bus 1050.

[0111] The CPU 1100 operates and controls each component based on programs stored in the ROM 1300 or the secondary storage device 1400. For example, the CPU 1100 loads the programs stored in the ROM 1300 or the secondary storage device 1400 into the RAM 1200 and executes processing corresponding to the various programs.

[0112] The ROM 1300 stores boot programs such as a Basic Input Output System (BIOS) executed by the CPU 1100 when the computer 1000 is started, and programs that depend on the hardware of the computer 1000 .

[0113] The secondary storage device 1400 is a computer-readable recording medium that non-temporarily records programs executed by the CPU 1100 and data used by such programs. Specifically, the secondary storage device 1400 is a recording medium that records at least the objective variable modification program according to this embodiment.

[0114] The communication interface 1500 is an interface for connecting the computer 1000 to an external network 1550. For example, the CPU 1100 receives data from other devices and transmits data generated by the CPU 1100 to other devices via the communication interface 1500.

[0115] The input / output interface 1600 is an interface for connecting the input / output device 1650 and the computer 1000. For example, the CPU 1100 receives data from an input device such as a keyboard or a mouse via the input / output interface 1600. The CPU 1100 also transmits data to an output device such as a display, a speaker, or a printer via the input / output interface 1600. The input / output interface 1600 may also function as a media interface for reading programs and the like recorded on a predetermined recording medium. Examples of media include optical recording media such as a DVD (Digital Versatile Disc) or a PD (Phase Change Rewritable Disk), magneto-optical recording media such as an MO (Magneto-Optical disk), tape media, magnetic recording media, and semiconductor memories.

[0116] For example, when the computer 1000 functions as the inspection apparatus 10, the CPU 1100 of the computer 1000 executes a program loaded onto the RAM 1200 to realize the functions of the control unit 12. The secondary storage device 1400 stores the objective variable correction program according to this embodiment and data in the storage unit 11. The CPU 1100 reads and executes program data 1450 from the secondary storage device 1400, but as another example, the CPU 1100 may obtain these programs from another device via an external network 1550.

[0117] <<7. Conclusion>> As described above, according to one embodiment of the present disclosure, the inspection device 10 includes a learning unit 12b, an inference unit 12ca, a weighting unit 12cb, and a correction unit 12cc. The learning unit 12b uses a teacher dataset 11a (corresponding to an example of a "dataset of teacher data including labels as target variables") to train a learning model 11b. The inference unit 12ca performs inference using the learning model 11b using the teacher dataset 11a. The weighting unit 12cb weights, by an arbitrary weight value η, a correction amount ω indicated by the difference between a first feature corresponding to each piece of teacher data in the feature space of the learning model 11b and a second feature obtained by the inference. The correction unit 12cc calculates the weighted correction amount ω. m This makes it possible to improve the inference accuracy of the learning model 11b while correcting mislabels, even when any data is used as training data.

[0118] Although the embodiments of the present disclosure have been described above, the technical scope of the present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present disclosure. Furthermore, components of different embodiments and modifications may be combined as appropriate.

[0119] Furthermore, the effects of each embodiment described in this specification are merely examples and are not intended to be limiting, and other effects may also be obtained.

[0120] The present technology can be configured as follows. (1) a learning unit that learns a learning model using a dataset of training data including labels as objective variables; an inference unit that performs inference using the learning model using the dataset of the training data; a weighting unit that weights, by an arbitrary weight value, a correction amount indicated by a difference between a first feature corresponding to each of the teacher data in a feature space of the learning model and a second feature obtained by the inference; a correction unit that corrects the label based on the weighted correction amount; An inspection device comprising: (2) a presentation unit that presents to a user the teacher data that is a candidate for correction, the label of which is corrected by the correction unit; The inspection device according to (1) further comprises: (3) the presentation unit presents to the user an interface that allows the user to specify the weight value; The inspection device according to (2) above. (4) when the user specifies a new weight value to the interface, the presentation unit causes the weighting unit to newly weight the correction amount using the specified weight value. The inspection device according to (3) above. (5) the presentation unit presents to the user the correction candidates in which the label is corrected based on the correction amount weighted by the weight value. The inspection device according to (2), (3) or (4) above. (6) the learning unit learns the learning model using the dataset of the training data whose labels have been corrected by the correction unit; the inference unit performs the inference using the learning model using the dataset of the training data whose labels have been corrected by the correction unit; the presentation unit presents to the user the inference results of the inference unit before and after the correction by the correction unit. The inspection device according to any one of (2) to (5) above. (7) The inference result includes at least a correct answer rate according to the learning model. The inspection device according to (6) above. (8) the correction unit determines whether or not the label can be corrected based on the weight value. The inspection device according to any one of (1) to (7) above. (9) The weight value is a weight in a weighted moving average. The inspection device according to any one of (1) to (8) above. (10) The inference unit performing the inference every time the learning unit performs learning using the data set of the teacher data a predetermined number of times; the weighting unit weights the correction amount each time the inference unit performs the inference; the correction unit corrects the label based on the weighted correction amount each time the weighting unit weights the correction amount; The inspection device according to (9) above. (11) The teacher data is image data obtained by visualizing sound. The inspection device according to any one of (1) to (10) above. (12) A computer-implemented method for modifying a target variable, comprising: a learning process of learning a learning model using a dataset of training data including labels as objective variables; an inference step of performing inference using the learning model using the dataset of the training data; a weighting step of weighting, by an arbitrary weight value, a correction amount indicated by a difference between a first feature corresponding to each of the teacher data in a feature space of the learning model and a second feature obtained by the inference; a modification step of modifying the label based on the weighted modification amount; Methods for correcting the target variable, including: (13) A learning procedure for learning a learning model using a dataset of training data including labels as a target variable; an inference procedure for performing inference using the learning model using the dataset of the training data; a weighting step of weighting, by an arbitrary weight value, a correction amount indicated by a difference between a first feature corresponding to each of the teacher data in a feature space of the learning model and a second feature obtained by the inference; a modification step of modifying the label based on the weighted modification amount; A target variable modification program that causes a computer to execute the above. [Explanation of symbols]

[0121] 3 Annotation device 5 UI device 7. Inspection data acquisition device 10 Inspection equipment 11 Storage section 11a Training dataset 11b Learning Model 11c Inference result information 12 Control Unit 12a Acquisition part 12b Learning Department 12c Teacher Data Correction Department 12d Presentation part 12e Inspection Department 51 Correction candidate display area 52 Inference accuracy display section

Claims

1. a learning unit that learns a learning model using a dataset of training data including labels as objective variables; an inference unit that performs inference using the learning model using the dataset of the training data; a weighting unit that weights, by an arbitrary weight value, a correction amount indicated by a difference between a first feature corresponding to each of the teacher data in a feature space of the learning model and a second feature obtained by the inference; a correction unit that corrects the label based on the weighted correction amount; An inspection device comprising:

2. a presentation unit that presents to a user the teacher data that is a candidate for correction, the label of which is to be corrected by the correction unit; The inspection device of claim 1 further comprising:

3. the presentation unit presents to the user an interface that allows the user to specify the weight value; The inspection device according to claim 2 .

4. When the user specifies a new weight value to the interface, the presentation unit causes the weighting unit to newly weight the correction amount using the specified weight value. The inspection device according to claim 3 .

5. the presentation unit presents to the user the correction candidate in which the label is corrected based on the correction amount weighted by the weight value. The inspection device according to claim 2 .

6. the learning unit learns the learning model using the dataset of the training data whose labels have been corrected by the correction unit; the inference unit performs the inference using the learning model using the dataset of the training data whose labels have been corrected by the correction unit; the presentation unit presents to the user the inference results of the inference unit before and after the correction by the correction unit. The inspection device according to claim 2 .

7. The inference result includes at least a correct answer rate according to the learning model. The inspection device according to claim 6.

8. the correction unit determines whether or not the label can be corrected based on the weight value. The inspection device according to claim 1 .

9. The weight value is a weight in a weighted moving average. The inspection device according to claim 1 .

10. The inference unit performing the inference every time the learning unit performs learning using the data set of the teacher data a predetermined number of times; the weighting unit weights the correction amount each time the inference unit performs the inference; the correction unit corrects the label based on the weighted correction amount each time the weighting unit weights the correction amount; The inspection device according to claim 9.

11. The teacher data is image data obtained by visualizing sound. The inspection device according to any one of claims 1 to 10.

12. A computer-implemented method for modifying a target variable, comprising: a learning process of learning a learning model using a dataset of training data including labels as objective variables; an inference step of performing inference using the learning model using the dataset of the training data; a weighting step of weighting, by an arbitrary weight value, a correction amount indicated by a difference between a first feature value corresponding to each of the teacher data in a feature space of the learning model and a second feature value obtained by the inference; a modification step of modifying the label based on the weighted modification amount; A method for correcting the target variable, including

13. A learning procedure for learning a learning model using a dataset of training data including labels as a target variable; an inference procedure for performing inference using the learning model using the dataset of the training data; a weighting step of weighting, by an arbitrary weight value, a correction amount indicated by a difference between a first feature corresponding to each of the teacher data in a feature space of the learning model and a second feature obtained by the inference; a modification step of modifying the label based on the weighted modification amount; A target variable modification program that causes a computer to execute the above.

Images